OSWorld 2.1 (offline subset, strict)多模态Agent

Pass@1 averaged over five attempts per task; 1080p, at most 500 actions, Claude Opus 4.8 grader. Separate from full-set and older offline-version results.OSWorld 2.1 official offline subset: 82 of 108 tasks, with no internet access. Tasks with every checkpoint satisfied.

排行数据源

#模型得分

1

Claude Sonnet 5.5Anthropic

48.8%

2

Claude Haiku 5.5Anthropic

37.1%

3

GPT-6 LunaOpenAI

17.1%

4

Claude Haiku 4.5Anthropic

1.7%

共 4 个模型

返回基准目录